Source-linked AI summary
Map Enhanced Route Travel Time Prediction using Deep Neural Networks
Soumi Das, Rajath Nandan Kalava, Kolli Kiran Kumar, Akhil Kandregula, Kalpam Suhaas, Sourangshu Bhattacharya, Niloy Ganguly
TL;DR
Travel-time prediction models based on trajectory segments have generally left road-network information out of the prediction process. This paper incorporates road-node embeddings and map distance into an existing deep-learning model, with improved accuracy reported especially when training data are limited or road and coordinate distances differ more.
Problem
Recent deep-learning travel-time models estimate times for trajectory segments but do not account for road-network information.
Source: §1 · p. 1Method
The paper incorporates road-node embeddings and source-to-destination map distance into an existing deep-learning travel-time model.
Source: §1 · p. 1, §3.3 · p. 2Results
Takeaways & Limitations
The reported results suggest road-network information provides additional information for predicting travel times.
Source: §1 · p. 1
Abstract
from arXiv · showhide
Travel time estimation is a fundamental problem in transportation science with extensive literature. The study of these techniques has intensified due to availability of many publicly available large trip datasets. Recently developed deep learning based models have improved the generality and performance and have focused on estimating times for individual sub-trajectories and aggregating them to predict the travel time of the entire trajectory. However, these techniques ignore the road network information. In this work, we propose and study techniques for incorporating road networks along with historical trips' data into travel time prediction. We incorporate both node embeddings as well as road distance into the existing model. Experiments on large real-world benchmark datasets suggest improved performance, especially when the train data is small. As expected, the proposed method performs better than the baseline when there is a larger difference between road distance and Vincenty distance between start and end points.
1 INTRODUCTION
The introduction frames travel-time prediction as an established problem and notes that a successful deep-learning approach does not use road-network information. The paper proposes incorporating mapped road nodes, node embeddings, and map distance into prediction.
1 INTRODUCTION
- Geo-convolution predicts travel times for short trajectory segments and stitches them together, but does not account for the road network.The paper investigates explicitly incorporating map information into prediction.Source: §1 · p. 1
- The proposed approach attributes trajectory points to road-network nodes and uses node embeddings and source-to-destination map distance.Source: §1 · p. 1
- The authors report improved overall accuracy, especially with fewer training trajectories and when actual and map distances differ more.They interpret the latter pattern as evidence that road-network information provides additional predictive information.Source: §1 · p. 1
2 RELATED WORK
Prior travel-time research ranges from static data collection to GPS-trace-based deep learning. The reviewed methods use neural networks to learn features and spatial or temporal patterns, including trajectory-segment representations.
2 RELATED WORK
- Recent work uses publicly collected GPS traces [12] to train deep-learning models for travel-time prediction.Earlier approaches mostly relied on static data collection techniques.Source: §2 · p. 1
- Siripanpornchana et al. use a Deep Belief Network with stacked Restricted Boltzmann Machines for unsupervised feature learning, while LSTMs have been used for sequential patterns and automatic feature selection.Source: §2 · p. 1
- Zhang et al. use BiLSTMs to learn spatio-temporal grid parameters with auxiliary supervision [12], while Wang et al. combine geo-convolution and stacked LSTMs.The latter architecture is intended to capture short-term segment signals and long-term trajectory prediction.Source: §2 · p. 1
3 TRAVEL TIME PREDICTION
This section defines trip travel-time prediction on mapped and unmapped trajectories, describes the baseline neural architecture, and presents embedding-based modifications that incorporate road-network information.
3.1 Problem Definition
- For a mapped trip S or unmapped trip T, the prediction task is to learn a function that outputs a positive travel-time estimate.The experiments use mapped trips.Source: §3.1 · p. 2
3.2 Node Geo-convolution model
- The baseline combines trip attributes, geo-convolution over successive nodes, and an RNN; attention pooling predicts whole-path time while training weights local- and whole-path losses.The convolution learns spatial features from local paths, and the RNN captures temporal correlations in those features.Source: §3.2 · p. 2
3.3 Embedding enhanced Geo-convolution model
- The proposed model learns Node2Vec representations for mapped road nodes and uses 128-dimensional embeddings to represent each node in a trip.The authors expect nearby nodes and edges to have similar embeddings, capturing spatial dependencies among roads.Source: §3.3 · p. 2
- Two modifications feed node embeddings into the baseline: E-GC uses embeddings alone, while EL-GC concatenates embeddings with mapped-node coordinates.Both retain the baseline convolution and recurrent-layer procedure.Source: §3.3 · pp. 2–3
4 EXPERIMENTAL RESULTS
The experiments compare map-enhanced travel-time models with the baseline on Porto and Beijing data, using coordinate, map, and combined distances. Both proposed modifications generally outperform the baseline, with performance patterns varying across distance buckets.
4 EXPERIMENTAL RESULTS
- The evaluation compares the baseline and proposed methods on Porto and Beijing test sets using mean absolute percentage error.Source: §4 · p. 3, §4.2 · p. 3
4.1 Dataset Creation
- The Porto and Beijing experiments use 326,424 and 15,772 trips, respectively, and map trip locations to OpenStreetMap road nodes using LSH.Mapping errors stay below 0.2 km for Porto and 0.25 km for Beijing; Porto data uses a selected 0.55 × 0.55 region.Source: §4.1 · p. 3, Figure 1 · p. 3
4.2 Results
- Both E-GC and EL-GC outperform L-GC using coordinate, map, and combined distances; the coordinate-distance margin is larger in Beijing than Porto.With map distance, the margin is lower; the combined-distance experiment concatenates map and coordinate distances to the convolution output.Source: §4.2 · p. 4
- EL-GC performs best overall across Beijing path lengths for both coordinate and map distances, while E-GC degrades on longer map-distance paths.L-GC has high MAPE on shorter coordinate-distance paths, which decreases as paths get longer.Source: §4.2 · p. 4
5 CONCLUSION
The paper modifies a deep-neural-network travel-time model to incorporate road-network information. The authors report that map distance and node embeddings outperform the baseline, and the figures describe MAPE across distance buckets and distance-gap buckets.
5 CONCLUSION
- The proposed modifications extend an existing state-of-the-art deep neural network for travel-time prediction.Source: §5 · p. 4
- The figures plot Beijing MAPE by coordinate/map distance buckets and by the difference between those distances.Source: Figure 2 · p. 4, Figure 3 · p. 4
- Map distance and node embeddings outperform the baseline, with embeddings described as capturing nearby-node relationships and spatial dependencies.Source: §5 · p. 4